The Narrative World Model (NWM) is a writer-memory system designed to handle complex narrative states by combining a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. This approach addresses the limitations of general-purpose systems that fail to capture the structural dependencies required for multi-hop questions in long-form fiction.
- NWM pairs a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval.
- Evaluation uses a held-constant Opus 4.8 reader on a reproducible public corpus and a validated multi-hop benchmark.
- The system is compared against Graphiti/Zep, the strongest existing temporal-knowledge-graph agent-memory framework.
- NWM substantially outperforms Graphiti/Zep, GraphRAG, and flat retrieval on multi-hop narratological QA.
- Performance gains are attributed to representational structure rather than extractor quality or graph size.
The authors consider this significant because the advantage is representational, surviving even when the baseline is rebuilt with NWM's own extractor, proving the value of narratology-grounded structures for narrative memory.